CAF-enriched COL4A6 promotes gastric cancer progression and stromal remodeling despite an inverse association in bulk transcriptomes

Abstract

The progression of gastric cancer (GC) is profoundly influenced by extracellular matrix (ECM) remodeling within the tumor microenvironment (TME); however, the compartment-specific roles of key ECM proteins (e.g., epithelium vs. stroma) remain incompletely understood. By integrating multi-omics bulk transcriptomic data with single-cell sequencing data, we observed a bulk-compartment discrepancy for COL4A6: COL4A6 carries an inverse association (OR < 1) in bulk transcriptome-based risk models, whereas its protein is markedly upregulated and predominantly localized to the CAF compartment in clinical specimens. We constructed a Matrisome Risk Score (MRS) model and identified ECM subtypes with distinct prognoses, confirming that the high-matrix subtype (MC1) is closely associated with immunosuppression and chemotherapy resistance. Functional experiments demonstrated that COL4A6 from CAF-enriched fibroblast cultures promotes tumor progression and stromal remodeling by inducing CAF activation (upregulation of α-SMA and FAP, remodeling of stress fibers) and promoting EMT in gastric cancer cells. Collectively, our findings support a compartment-dependent model in which epithelial silencing/loss may bias bulk mRNA associations, whereas CAF-compartment COL4A6 accumulation is a functional driver of GC progression. This highlights CAF-compartment COL4A6 as a potential target for therapies aimed at ECM remodeling within the TME.

Supplementary Information

The online version contains supplementary material available at 10.1038/s41598-026-38120-7.

Introduction

Gastric Cancer (GC) ranks as the fifth most common malignancy worldwide and one of the leading causes of cancer-related mortality. Owing to its insidious onset, high aggressiveness, and propensity for early metastasis, the prognosis for GC patients remains exceptionally poor¹. Although comprehensive treatment strategies centered on surgery—supplemented by chemotherapy, radiotherapy, and emerging immunotherapies—have made progress in recent decades^2,3, the 5-year survival rate for GC remains unsatisfactory. This is largely attributed to the tumor’s high heterogeneity, complex genetic background, and prevalent drug resistance mechanisms, posing a significant burden on global public health⁴.

Growing evidence indicates that GC pathogenesis is a complex, multi-step process dictated not only by the intrinsic properties of cancer cells (e.g., genetic mutations, epigenetic alterations) but also profoundly regulated by the complex Tumor Microenvironment (TME) it inhabits^5,6. The TME is a dynamic and heterogeneous ecosystem composed of various cellular components (such as stromal cells, immune cells, and endothelial cells) and non-cellular components, namely the Extracellular Matrix (ECM)⁷. The ECM serves as more than just a physical scaffold for tissue structure; it is a bioactive reservoir containing collagens, proteoglycans, glycoproteins, and diverse signaling molecules⁸. Interactions within the TME, particularly those mediated by ECM components, profoundly influence cancer cell communication, proliferation, survival, migration, and therapeutic resistance by activating complex networks like the Integrin signaling pathway. These interactions are critical factors driving tumor progression, invasion, and metastasis^9–15.

However, the roles of ECM components in tumors are highly context- and compartment-dependent, varying by cellular source, microenvironmental state, and analytical resolution. Take COL4A6, a key subunit of Type IV collagen (the main component of the basement membrane (BM)), as an example. Its role in cancer is highly debated in the scientific community. On one hand, the BM acts as the first physical barrier for epithelial cells. Several studies (including in colorectal¹⁶ and prostate cancer) have reported that COL4A6 and its adjacent gene, COL4A5, are often co-silenced in epithelial cells due to promoter hypermethylation. The loss of their expression is thought to disrupt BM integrity and is associated with stronger invasiveness and poor prognosis¹⁷, aligning with a classic "tumor-suppressive/protective" profile.

On the other hand, during TME remodeling, stromal cells, predominantly Cancer-Associated Fibroblasts (CAFs), secrete and deposit copious amounts of ECM proteins (including collagens), leading to matrix fibrosis and increased stiffness. In this context, high expression of COL4A6 has been shown in some cancers (such as ovarian and liver cancer) to promote invasion and drug resistance via the DDR1/Integrin signaling pathway, thus presenting a "pro-tumorigenic" phenotype^18,19.

In gastric cancer (STAD), bulk-level transcriptomic associations for COL4A6 differ across cohorts and analytical frameworks, with some studies implicating epithelial downregulation while others link higher COL4A6 to adverse features. Such differences are consistent with cell-type composition effects and compartment-specific expression, and they motivate a compartment-resolved assessment of COL4A6 localization and function in GC.

Therefore, clarifying the compartment-dependent localization and function of COL4A6 in GC is important for understanding ECM remodeling in the TME. This study (Supplementary Fig. 1) aims to integrate TCGA transcriptomic subtyping, single-cell data analysis, and functional validation. We first seek to depict the macroscopic ECM landscape of GC to identify clinically significant matrisome subtypes. Subsequently, we focus at the microscopic level on the bulk-compartment discordance of COL4A6 to investigate the mechanisms by which CAF-enriched COL4A6 contributes to GC progression, ECM remodeling, and CAF activation.

Methods

Data acquisition

Transcriptomic sequencing data for 374 tumor tissues (from 343 patients) and adjacent normal tissues (from 31 patients) for Gastric Adenocarcinoma (STAD) were downloaded using the TCGAbiolinks package²⁰. The validation dataset, GSE66229, comprised microarray data from 400 samples, including tumor tissues (n = 300) and adjacent normal tissues (n = 100)¹⁶. All patients in GSE66229 had not received neoadjuvant chemotherapy or radiotherapy; the platform was the Affymetrix Human Genome U133 Plus 2.0 Array. Samples from GSE66229 were sourced from the Samsung Medical Center, Korea. The single-cell dataset, GSE183904, contained 40 scRNA-seq samples, including 26 GC tumor samples, 10 adjacent normal samples, 3 peritoneal metastasis samples, and 1 normal peritoneum sample²¹. All GSE183904 samples were from the National University Hospital, Singapore, and sequenced on the Illumina NovaSeq 6000 platform. For this study, 36 GC tumor and adjacent normal samples from GSE183904 were selected for scRNA-seq data analysis. The definition of the “Matrisome” was primarily derived from the database established by Naba et al.⁸, including 1068 core matrisome genes and 753 matrisome-associated genes.

Data processing

Differential analysis of the TCGA-STAD raw count data was performed using the negative binomial distribution method within the DESeq2 package. The criteria for screening differentially expressed genes (DEGs) were |logFC|> 1 and adjusted P < 0.05. Data in transcripts per million (TPM) format were used for Principal Component Analysis (PCA) and molecular subtyping analysis. GSE66229 data were downloaded via the GEOquery package, log2-transformed, and batch-corrected using the limma package²² for subsequent analysis.

Construction of the matrisome risk score (MRS)

Risk feature screening was performed on the differential matrisome genes from TCGA-STAD. First, LASSO regression was employed using the minimum Lambda value criterion to eliminate collinear features. The Random Forest algorithm was then used to rank gene importance based on the "Mean Decrease Gini," screening the top 30 key genes. Finally, SVM-RFE was utilized to select the optimal feature subset with the highest classification accuracy. Finally, we took the intersection of the candidate gene sets selected by all three machine learning methods to obtain the final core genes. The glm function in the R stats package was used to calculate the odds ratio (OR) for each core gene. The Matrisome Risk Score (MRS) was calculated using the following formula:

(where Genei represents the z-score normalized value of gene i, and ORi represents the odds ratio of gene i).

Subtype identification

The M3C package²³ was used to perform K-means consensus clustering on the differential matrisome genes. The optimal K value was determined based on the maximum Relative Cluster Stability Index (RCSI) and P < 0.05 to differentiate sample subtypes.

Pathway enrichment and immune infiltration analysis

The clusterProfiler package²⁴ was used for GO and KEGG^25,26 analysis (enrichment criteria: P < 0.05, BH-corrected). Hallmark pathways and C2 modules were downloaded from the MSigDB database²⁷. The ssgsea algorithm in the IOBR package²⁸ was used to calculate pathway enrichment levels, and the limma package²² was used to extract differentially enriched pathways. The ESTIMATE²⁹ and CIBERSORT³⁰ algorithms within IOBR were used to assess immune/stromal scores and immune cell infiltration abundance.

Ligand-receptor interaction analysis

Based on the dataset from Ramilowski et al.³¹, receptor-ligand pairs where the core matrisome genes acted as ligands were extracted for analysis. Interaction strength was defined as the product of the normalized expression values of the ligand and its receptor.

Cell component deconvolution of bulk RNA-seq

Three methods—ESTIMATE, CIBERSORT, and the MuSiC package³² (combined with GSE183904 single-cell data)—were used to evaluate the cellular composition of TCGA-STAD samples (stromal score, immune score, tumor purity, and various cell abundances).

Single-cell dataset processing and cell communication

The UMI count matrix from GSE183904 was processed in R (v4.2.1; https://www.r-project.org/) using Seurat (v4.0.6; https://satijalab.org/seurat/) following a standard workflow, including quality control, normalization, dimensionality reduction, and clustering. Cell–cell communication analysis was performed using the CellChat package (v1.6.1; https://github.com/sqjin/CellChat)³³. Communication networks were computed and visualized using the aggregateNet, netAnalysis_computeCentrality, netVisual_circle, and netVisual_bubble functions. For cell-type-resolved validation of COL4A6, quantitative analyses were conducted at the major-compartment level, encompassing all annotated cellular compartments (including epithelial cells, fibroblasts/CAFs, immune cells, endothelial cells, and pericytes). COL4A6 expression across cell types was visualized using Seurat functions FeaturePlot and VlnPlot. Cells with a log-normalized COL4A6 expression value greater than 0 were defined as COL4A6-positive, and the fraction of COL4A6-positive cells was calculated for each major compartment.

Patients and tissue specimens

All 15 pairs of GC and adjacent normal tissue specimens were collected at Shanghai Changzheng Hospital between 2018 and 2022. No patients had received neoadjuvant chemoradiotherapy, and all provided written informed consent. This study was approved by the Medical Ethics Committee of Shanghai Changzheng Hospital (2022SL024).

Primary CAF culture

CAF-enriched fibroblast cultures used for functional assays were obtained from patient-derived gastric cancer tissues. Fresh tumor tissues were minced into ~ 1 mm³ fragments and enzymatically digested (collagenase and DNase I) at 37 °C with agitation. Cell suspensions were filtered through a 70-µm strainer, centrifuged, and plated in DMEM with 10% FBS. After 12–24 h, non-adherent cells were removed by medium replacement, and adherent spindle-shaped fibroblast-like cells were further expanded. Cells were maintained in DMEM supplemented with 10% FBS at 37 °C and 5% CO₂. To minimize culture drift, all experiments were performed using early-passage CAF-enriched fibroblast cultures (P2-P5). The CAF-like phenotype was supported by expression of canonical CAF markers FAP and α-SMA as shown in our immunofluorescence and western blot analyses (Fig. 6C–F). We acknowledge that additional negative lineage markers (e.g., EPCAM, CD31, CD45) were not quantitatively assessed in this study and therefore interpret the functional results as derived from CAF-enriched fibroblast cultures. Primary CAF cultures were established from n = 2 independent gastric cancer patients. Unless otherwise stated, key functional experiments were repeated using at least two independent CAF preparations, and we observed consistent qualitative trends across the two independent CAF preparations. Accordingly, throughout the manuscript we use the term “CAF-enriched fibroblast cultures” to reflect the enrichment-based nature of the preparation, and we interpret functional effects as attributable to fibroblast/CAF-like stromal cells rather than purified CAFs.

Fig. 6.

COL4A6 induces CAF activation and functional cytoskeletal reorganization. (A) and (B) Phalloidin staining (red) of CAFs. Shows disorganized cytoskeleton in the sh-COL4A6 group (A), whereas the OE-COL4A6 group (B) forms thick, parallel F-actin stress fibers. DAPI (blue) for nuclear staining. (C) and (D) Immunofluorescence co-staining of CAFs. Shows reduced expression of FAP (green) and α-SMA (red, activation marker) in the sh-COL4A6 group (C), and significantly enhanced FAP and α-SMA expression in the OE-COL4A6 group (D). (E) and (F) Western Blot analysis of CAFs. Confirms downregulated α-SMA and FAP protein levels in the sh-COL4A6 group (E), and upregulated α-SMA and FAP protein levels in the OE-COL4A6 group (F). (*P < 0.05, **P < 0.01, ***P < 0.001; ns, not significant).

Cell culture and lentiviral transduction

HGC27 and AGS cells were purchased from the Cell Bank of the Chinese Academy of Sciences (Shanghai, China) and cultured in RPMI-1640 medium (Wisent, 350-000-CL) supplemented with 10% fetal bovine serum (FBS; Gibco, 10100147C) and 1% penicillin–streptomycin at 37 °C in a 5% CO₂ incubator. To modulate COL4A6 expression in CAFs, we constructed stable lentiviral transduction lines. Short hairpin RNA (shRNA) sequences targeting COL4A6 (see Supplementary Table S1) and the full-length COL4A6 overexpression sequence (OE-COL4A6) were cloned into lentiviral vectors (GenePharma, Shanghai, China). Empty vectors served as negative controls (sh-NC and OE-NC, respectively). The lentiviral packaging and infection procedures were as follows: lentiviral vectors and packaging plasmids (pMD2.G and psPAX2) were co-transfected into HEK293T cells using Lipofectamine 3000 (Invitrogen). After 48 h, the virus-containing supernatant was collected and filtered through a 0.45 μm filter. CAF-enriched fibroblast cultures were seeded in six-well plates and infected with the viral supernatant supplemented with 8 μg/mL Polybrene (Sigma-Aldrich). Forty-eight hours post-infection, cells were continuously selected with 2 μg/mL Puromycin (Sigma-Aldrich) for one week to obtain stably transfected cell lines. Transfection efficiency was verified by qRT-PCR and Western Blot.

Co-culture system

A schematic of the CAF-GC cell co-culture system is shown in Fig. 4B. GC cells (1 × 10^5) and transfected CAFs (1 × 10^5) were seeded into the lower and upper chambers, respectively, of a six-well Transwell apparatus (Corning Incorporated) with 0.4 µm pores. After one week of co-culture, the GC cells in the lower chamber were collected for subsequent functional experiments.

Fig. 4.

CAF-compartment-associated COL4A6 promotes migration, invasion, and EMT of GC cells. (A) Western Blot validation of COL4A6 protein expression following regulation by shRNA (sh-COL4A6) and overexpression plasmid (OE-COL4A6) in CAFs. (B) Schematic of the Transwell co-culture system for GC cells and CAFs. (C) Wound healing (upper panel) and Transwell invasion (lower panel) assays for the AGS cell line. (D) Wound healing (upper panel) and Transwell invasion (lower panel) assays for the HGC27 cell line. (Quantitative statistics for C and D on the right). (E) Western Blot analysis of EMT markers (E-cadherin, N-cadherin, Vimentin) for the AGS cell line. (F) Western Blot analysis of EMT markers (E-cadherin, N-cadherin, Vimentin) for the HGC27 cell line. (Quantitative statistics for E and F on the right). (*P < 0.05, **P < 0.01, ***P < 0.001; ns, not significant).

qRT-PCR

Total RNA was extracted from cultured cells or ground tissues using TRIzol reagent (Invitrogen, USA) according to the manufacturer’s instructions. RNA concentration and purity were determined using a NanoDrop 2000 spectrophotometer (Thermo Scientific). Subsequently, 1 μg of total RNA was reverse-transcribed into cDNA using HiScript III All-in-one RT SuperMix (Vazyme, Nanjing, China, R333-01). Real-time quantitative PCR (qRT-PCR) was performed on a QuantStudio 7 Flex System (Thermo Fisher Scientific) using AceQ qPCR SYBR Green Master Mix (Vazyme, Q141-02). The amplification program was as follows: pre-denaturation at 95 °C for 5 min; followed by 40 cycles of 95 °C for 10 s and 60 °C for 30 s. GAPDH was used as the internal control gene, and the relative expression of COL4A6 was calculated using the 2^-ΔΔCt method. Primer sequences are detailed in Supplementary Table S2.

Western blotting (WB)

Cells were harvested and lysed on ice for 30 min using RIPA lysis buffer (Beyotime, China) containing protease and phosphatase inhibitors (Roche). The supernatant was collected by centrifugation, and protein concentration was determined using a BCA Protein Assay Kit (Thermo Scientific). Equal amounts of protein (20–30 μg per lane) were separated by 8–10% SDS-PAGE gel electrophoresis and subsequently transferred onto polyvinylidene fluoride (PVDF) membranes (Millipore, USA). The membranes were blocked with TBST buffer containing 5% non-fat milk for 1 h at room temperature and then incubated with specific primary antibodies overnight at 4 °C. Primary antibodies included: Anti-COL4A6 (Abcam, dilution 1:1000), Anti-E-cadherin (CST, 1:1000), Anti-N-cadherin (CST, 1:1000), Anti-Vimentin (CST, 1:1000), Anti-α-SMA (Abcam, 1:1000), Anti-FAP (Abcam, 1:1000), and Anti-β-actin (Sigma, 1:5000) (see Supplementary Table S3 for details). The following day, membranes were washed three times with TBST and incubated with HRP-conjugated secondary antibodies (Goat anti-rabbit/mouse IgG, 1:5000, Jackson ImmunoResearch) for 1 h at room temperature. Finally, bands were visualized using an ECL chemiluminescence kit (Thermo Scientific) and detected via a gel imaging system (Bio-Rad). Grayscale quantitative analysis was performed using ImageJ software.

Immunohistochemistry (IHC)

Tissue samples were fixed in 10% neutral formalin, embedded in paraffin, and cut into 4 μm thick sections. Sections were deparaffinized in xylene and rehydrated in graded ethanol. For antigen retrieval, sections were immersed in citrate buffer (pH 6.0), boiled in a microwave for 15 min, and allowed to cool naturally to room temperature. Endogenous peroxidase activity was blocked by incubating with 3% hydrogen peroxide (H₂O₂) solution for 10 min at room temperature, and nonspecific binding was blocked using 5% bovine serum albumin (BSA) for 30 min. Sections were incubated with rabbit monoclonal anti-COL4A6 antibody (Abcam, 1:200 dilution) overnight in a humidified chamber at 4 °C. The next day, sections were incubated with HRP-labeled secondary antibodies for 1 h at room temperature, stained with DAB (3,3'-diaminobenzidine) chromogen, and counterstained with hematoxylin. Staining results were evaluated by two experienced pathologists under double-blind conditions. The IHC score was calculated by multiplying the staining intensity (0 = no staining, 1 = weak, 2 = moderate, 3 = strong) by the percentage of positive cells (0 = 0–5%, 1 = 6–25%, 2 = 26–50%, 3 = 51–75%, 4 = 76–100%), resulting in a total score ranging from 0 to 12. A total score > 3 was defined as high COL4A6 expression.

Animal experiments

Four-week-old female BALB/C nude mice were purchased from GemPharmatech. All methods were carried out in accordance with relevant guidelines and regulations. Experiments were approved by the Animal Experiment Ethics Committee of Shanghai Changzheng Hospital (2024SL014). All methods are reported in accordance with ARRIVE guidelines. For the subcutaneous xenograft model, 1.5 × 10⁶ HGC27 cells were mixed with 0.5 × 10⁶ stably transfected CAFs (sh-NC, sh-COL4A6, OE-NC, OE-COL4A6 groups) and injected into the axilla of the forelimb of 6-week-old nude mice (n = 3 per group). Tumor volume (V = π/6 (L × W²)) was measured every 3 days. Detailed individual tumor measurements and animal weights at each time point are provided in Supplementary Table S4. Four weeks after cell injection, mice were euthanized by CO2 asphyxiation using a gradual displacement rate (30% of the chamber volume per minute), followed by cervical dislocation as a secondary physical method, and tumors were collected and weighed.

Statistical analysis

SPSS 25.0 software was used, and data are presented as mean ± standard deviation (SD). Inter-group comparisons were performed using One-way ANOVA and LSD post hoc test. P < 0.05 was defined as statistically significant.

Results

The ECM landscape in the GC TME: systematic transcriptomic dysregulation

To systematically analyze the role of the ECM in the GC TME, we first performed a differential expression analysis of ECM genes (core matrisome and matrisome-associated) in the TCGA-STAD cohort, based on the classification by Naba et al.⁸ (Study workflow in Supplementary Fig. 1). The results revealed significant ECM remodeling between tumor and adjacent normal tissues: 45.5% of the 1068 core matrisome genes and 42.2% of the 753 matrisome-associated genes were differentially expressed. These differentially expressed ECM genes collectively accounted for 9.5% of all DEGs (Supplementary Fig. 2B, 2C). PCA further confirmed that the expression profile of these differential core matrisome genes alone could clearly distinguish tumor samples (red) from normal samples (blue) at the transcriptomic level (Supplementary Fig. 2A), highlighting the prevalence and importance of ECM remodeling in GC pathogenesis.

To investigate the heterogeneity of these differential genes, we conducted a correlation clustering analysis (Supplementary Fig. 2D). The results revealed three distinct gene clusters (Cluster 1–3) that exhibited starkly different cell-of-origin specificity in the GSE183904 single-cell data (Supplementary Fig. 2E):

• Cluster 1 (e.g., COL1A1, COL3A1, COL1A2, COL4A1): Primarily expressed in fibroblasts, endothelial cells, and pericytes, and significantly upregulated in tumor tissues.

• Cluster 2 (e.g., SPARCL1, FBLN1): Similar cell origins to Cluster 1, but with higher expression in adjacent normal tissues.

• Cluster 3: Broad cell origins, but with specific high expression in endothelial cells and pericytes in tumor tissue, and low expression in fibroblasts in adjacent normal tissue.

• These findings confirm that ECM dysregulation in the GC TME is highly heterogeneous and cell-specific, suggesting that ECM genes from different cellular origins may play distinct biological roles.

Screening core matrisome genes and constructing a prognostic risk model

To screen for the most critical drivers among the complex ECM genes, we employed three independent machine learning algorithms (LASSO regression, RF, SVM) for feature selection on the differential core matrisome genes (Supplementary Fig. 3A–E). By analyzing the intersection of the candidate gene sets selected by the three algorithms, we ultimately locked in four core matrisome genes: COL4A1, BGN, LAMC2, and COL4A6 (Fig. 1A).

Fig. 1.

Construction and diagnostic/prognostic value of the Matrisome Risk Score (MRS). (A) Odds ratios (OR) and 95% confidence intervals (CI) for the four core genes (BGN, COL4A1, LAMC2, COL4A6) selected by the intersection of the three models. (B) Receiver Operating Characteristic (ROC) curve for the MRS model in the TCGA-STAD training set (AUC = 0.958). (C) Comparison of MRS scores between tumor and normal samples in the TCGA-STAD training set. (D) Comparison of MRS scores between tumor and normal samples in the GSE66229 validation set. (E) Comparison of MRS scores between different age groups (old vs young). (F) Comparison of MRS scores between different TNM stages (stage III/IV vs stage I/II). (G) ROC curve for the MRS model in the GSE66229 validation set (AUC = 0.908).

Based on these four genes, we constructed a Matrisome Risk Score (MRS) model and calculated the odds ratio (OR) for each gene. Interestingly, COL4A6 showed an inverse association in this bulk transcriptome-based model (OR < 1), in stark contrast to the other selected matrisome genes—BGN, COL4A1, and LAMC2—which, consistent with the typical role of collagen family members as risk factors in gastric cancer (OR > 1), were identified as potent risk factors with ORs of 4.856, 4.467, and 3.198, respectively. In contrast, COL4A6 was the sole gene with an OR significantly less than 1 (OR = 0.516, 95% CI 0.397–0.661) (Fig. 1A), suggesting that it acts as a potential “protective” factor at the bulk tissue transcriptomic level in TCGA.

Despite this inverse association of COL4A6, the MRS model (including COL4A6 with OR < 1) still demonstrated robust diagnostic efficacy. In the TCGA training set, the MRS score efficiently distinguished tumor from normal tissues (AUC = 0.958, Fig. 1B), which was validated in the GSE66229 set (AUC = 0.908, Fig. 1G). The MRS score was significantly higher in tumor tissues than in normal tissues (Fig. 1C,D) and was independent of patient age (Fig. 1E), but was significantly positively correlated with clinical stage (III/IV vs. I/II) (Fig. 1F). This indicates that the overall extent of ECM remodeling is closely related to tumor progression.

ECM-based molecular subtypes of GC and their clinicopathological features

To understand the macroscopic impact of ECM heterogeneity on patient phenotypes, we performed consensus clustering based on the differential core matrisome genes, dividing TCGA-STAD patients into two distinct Matrisome Subtypes: MC1 and MC2 (Fig. 2A).

• MC1 (High-Matrisome Subtype): Characterized by high expression of ECM genes (especially Cluster 1 genes) (red region in heatmap).

• MC2 (Low-Matrisome Subtype): Characterized by low expression of ECM genes (blue region in heatmap).

Fig. 2.

ECM-based molecular subtypes (MC1/MC2) correlate with GC prognosis. (A) Consensus clustering heatmap of differential core matrisome genes, showing two subtypes (MC1, MC2) and corresponding clinical information (Sex, Age, Stage, Cluster). (B) Comparison of “KEGG_ECM_RECEPTOR_INTERACTION” pathway enrichment scores between MC1 and MC2 subtypes^25,26. (C) Comparison of “BIOCARTA_ECM_PATHWAY” pathway enrichment scores between MC1 and MC2 subtypes. (D) Kaplan–Meier curve of overall survival (Survival probability) for MC1 (blue) and MC2 (red) subtype patients in the TCGA-STAD cohort. (E) Kaplan–Meier curve of overall survival for MC1 (blue) and MC2 (red) subtype patients in the GSE66229 validation cohort.

Pathway enrichment analysis confirmed that “KEGG_ECM_RECEPTOR_INTERACTION” and “BIOCARTA_ECM_PATHWAY” were significantly higher in the MC1 subtype than in MC2 (Fig. 2B,C). Critically, Kaplan–Meier survival analysis showed that patients in the MC1 (high-matrisome) subtype had significantly shorter overall survival (OS) than those in the MC2 subtype (P = 0.014, Fig. 2D). This result was consistently observed in the GSE66229 validation cohort (P = 0.015, Fig. 2E), providing strong evidence that ECM remodeling is a key driver of poor prognosis in GC.

The unique TME cellular ecosystems of ECM subtypes

We further explored the differences in TME cellular composition between these two subtypes. The ESTIMATE algorithm showed that the MC1 (high-matrisome) subtype had significantly higher Immune and Stromal Scores, as well as lower Tumor Purity (Supplementary Fig. 4A–C), indicating denser infiltration of TME components (stromal and immune cells) and a relatively lower proportion of cancer cells.

To resolve the specific cell types, we combined CIBERSORT and MuSiC (single-cell deconvolution) for analysis (Supplementary Fig. 4D–I). The results depicted two distinct TME landscapes:

• MC1 (Cold Tumor / Immunosuppressive): TME enriched with fibroblasts, myeloid cells (Supplementary Fig. 4E,F), as well as endothelial cells and pericytes (Supplementary Fig. 4G,H). Among immune subpopulations, MC1 was significantly enriched with M2 macrophages, resting mast cells, and resting dendritic cells (Supplementary Fig. 4I). Although the MC1 subtype exhibited a high total immune score, it was predominantly dominated by immunosuppressive cells and lacked effector T cells. Consequently, we functionally defined this subtype as an “immunosuppressive” microenvironment, representing a typical “cold tumor”.

• MC2 (Hot Tumor / Immune-Activated): TME enriched with tumor epithelial cells (Supplementary Fig. 4D), consistent with high tumor purity. Among immune subpopulations, MC2 was significantly enriched with activated CD4 memory T cells, plasma cells, and activated dendritic cells (Supplementary Fig. 4I). This represents an immune-cell-infiltrated microenvironment, i.e., a "hot tumor."

Functional pathways and therapeutic sensitivity differences in ECM subtypes

To investigate the biological functions underlying these subtype differences, we performed Hallmark pathway enrichment analysis (Supplementary Fig. 5A). The MC1 (high-matrix/poor prognosis) subtype was significantly enriched in pathways related to invasion, metastasis, and TME remodeling, such as Epithelial-Mesenchymal Transition (EMT), Angiogenesis, TGF-β signaling, and Hypoxia (Supplementary Fig. 5D–G). Specifically, compared to the MC2 subtype, MC1 exhibited significantly higher enrichment scores for EMT, KRAS signaling (up), and TGF-β signaling pathways, providing molecular support for its more aggressive invasive characteristics.

Conversely, the MC2 (low-matrisome / good prognosis) subtype was enriched in pathways related to cell proliferation and core metabolism, such as G2M checkpoint, oxidative phosphorylation, and DNA repair (Supplementary Fig. 5A). The enrichment of the “DNA repair” pathway strongly suggested that the MC2 subtype might be more sensitive to platinum-based chemotherapy, which relies on DNA damage³⁴. To validate this hypothesis, we utilized the oncoPredict R package—a tool designed to predict in vivo drug response (IC50) based on gene expression profiles³⁵—to assess sensitivity to oxaliplatin, a first-line chemotherapeutic agent for gastric cancer. The results confirmed that the predicted half-maximal inhibitory concentration (IC50) for oxaliplatin was significantly lower in the MC2 subtype than in MC1 (Supplementary Fig. 5B,C), indicating that MC2 patients are more likely to benefit from platinum-based chemotherapy, which partially explains their better prognosis.

Finally, we analyzed the cell communication networks within the TME. Both ligand-receptor analysis (Supplementary Fig. 5H) and single-cell communication analysis (Supplementary Fig. 6A–H) showed that cellular interactions in the TME (especially among fibroblasts, endothelial cells, pericytes, and epithelial cells) were primarily mediated by Integrins and their ligands, such as LAMININ (Supplementary Fig. 6C,D).

Linking an inverse bulk association of COL4A6 to CAF-compartment localization and function

Our macroscopic analysis confirmed that the "high-matrisome (MC1)" phenotype is associated with poor prognosis, immunosuppression, and chemotherapy resistance. Notably, COL4A6 exhibited an inverse association (OR < 1) in the bulk transcriptome-based model (Fig. 1A), prompting us to examine its cellular source and spatial localization to interpret this bulk-level signal within a compartment-resolved framework.

To reconcile this bulk association with compartment-specific biology, we next examined COL4A6 protein expression and spatial localization in clinical specimens. IHC staining confirmed that, compared to its extremely weak expression in normal tissues (N1-N3), COL4A6 protein was significantly upregulated in tumor tissues (T1-T3) (Fig. 3A). Masson’s trichrome staining also visually demonstrated that from normal tissue (N) to tumor tissues of different stages (T1-T4), collagen deposition (blue) progressively increased, and ECM remodeling intensified (Fig. 3B).

Fig. 3.

Upregulation of COL4A6 in GC tissues and its spatial localization to the CAF compartment. (A) IHC staining for COL4A6. Shows low expression in normal tissues (N1, N2, N3) and high expression in tumor tissues (T1, T2, T3) (statistical analysis of IHC scores on the right). (B) Masson’s trichrome staining. Shows progressively increasing collagen deposition (blue) and ECM remodeling from normal tissue (N) to tumor tissues (T1-T4). (C) Immunofluorescence co-staining. Shows significant co-localization (yellow/orange signals in Merge image) of FAP (green, CAF marker) and COL4A6 (red) in tumor tissues (T1, T2). (*P < 0.05, **P < 0.01, ***P < 0.001; ns, not significant).

Crucially, immunofluorescence (IF) co-staining provided a key clue to explaining this bulk-compartment discordance: In tumor tissues (T1, T2), the COL4A6 signal (red) primarily co-localized with FAP (green), a specific marker for CAFs. This appeared as yellow/orange signals in the merged images (Fig. 3C).

Consistent with the IF co-localization, analysis of the GSE183904 scRNA-seq cohort showed that COL4A6 transcripts were enriched in fibroblast/CAF populations, with the highest average expression and the largest fraction of COL4A6-positive cells compared with epithelial, immune, and endothelial compartments (20.3% COL4A6 + CAF; Supplementary Figure S7A–S7C).

Together, these observations suggest that the inverse association (OR < 1) observed in bulk TCGA models may reflect epithelial downregulation and/or sample-level cellular composition effects, whereas the increased COL4A6 protein signal in our specimens is predominantly localized to the CAF-stromal compartment.

CAF-compartment COL4A6 drives GC progression and stromal remodeling

Based on the finding in Fig. 3C, we hypothesized that CAF-compartment-associated COL4A6 possesses a strong "pro-tumorigenic" function. We designed a series of functional experiments to validate this hypothesis.

First, we modulated COL4A6 expression in CAF-enriched fibroblast cultures (Fig. 4A) and co-cultured them with GC cells (AGS, HGC27) in a Transwell system (Fig. 4B). Wound healing and Transwell invasion assays showed that after co-culture with CAFs overexpressing COL4A6 (OE-COL4A6), the migration (Fig. 4C,D upper panels) and invasion (Fig. 4C,D lower panels) capabilities of GC cells were significantly enhanced. Conversely, knocking down COL4A6 in CAFs (sh-COL4A6) significantly inhibited GC cell migration and invasion.

Mechanistically, WB analysis showed that OE-COL4A6 CAFs induced EMT in GC cells, characterized by the downregulation of the epithelial marker E-cadherin and the upregulation of mesenchymal markers N-cadherin and Vimentin (Fig. 4E,F).

Next, we validated the in vivo function of COL4A6 using a subcutaneous co-injection model (HGC27 cells + transfected CAFs) (Fig. 5A). The results showed that tumors formed by the OE-COL4A6 CAF group had significantly larger volumes and weights than the control group, whereas tumor growth in the sh-COL4A6 CAF group was inhibited.

Fig. 5.

CAF-compartment-associated COL4A6 promotes tumor growth and ECM remodeling in vivo. (A) Subcutaneous xenograft model. Left: Gross specimens of solid tumors formed by co-injection of HGC27 cells with differently treated CAFs. Middle: Tumor Weight statistics. Right: Tumor Volume growth curve over time. (B) Masson’s trichrome staining of tumor tissues, showing significantly increased collagen deposition (blue) in the OE-COL4A6 group. (C) Immunofluorescence co-staining of tumor tissues, showing significantly enhanced FAP (green) and COL4A6 (red) signals and co-localization (Merge) in the OE-COL4A6 group. (*P < 0.05, **P < 0.01, ***P < 0.001; ns, not significant).

Histological analysis confirmed that tumors in the OE-COL4A6 group exhibited denser collagen deposition (Masson staining, Fig. 5B) and stronger FAP-COL4A6 co-localization signals (Fig. 5C), demonstrating that CAF-compartment-associated COL4A6 drives ECM remodeling in vivo.

COL4A6 functions by inducing CAF activation and cytoskeletal reorganization

Finally, we investigated how COL4A6 affects the CAFs themselves. We stained the CAFs from the co-culture system. Phalloidin staining (Fig. 6A,B) revealed that OE-COL4A6 did not “disrupt” the cytoskeleton, but rather induced significant cytoskeletal reorganization, forming numerous thick, parallel F-actin stress fibers. This was in stark contrast to the diffuse, disorganized F-actin morphology in the sh-COL4A6 group.

The formation of stress fibers is a hallmark of CAF activation into myofibroblasts. To confirm this, we examined activation markers. Both immunofluorescence (Fig. 6C,D) and Western Blot (Fig. 6E,F) confirmed that the protein expression levels of α-SMA and FAP were significantly upregulated in the OE-COL4A6 group CAFs.

Together, these data indicate that COL4A6 (likely via autocrine signaling) is a key inducer of CAF activation. It drives CAFs to undergo a phenotypic switch, reorganize their cytoskeleton, and thus acquire enhanced capabilities for stromal remodeling and promoting tumorigenesis.

Discussion

By integrating multi-omic analysis and functional experiments, this study systematically characterized the ECM landscape in the GC TME and, centrally, clarified the compartment-dependent localization and function of COL4A6 in gastric cancer and interpreted its inverse association in bulk transcriptome-based models.

The first core finding of this study is the dissection of an apparent bulk-to-compartment discordance for COL4A6. The function of COL4A6 in tumor biology has been reported to be context-dependent across cancer types^17–19. In our study, we observed a consistent bulk-to-compartment discordance: COL4A6 showed an inverse association in bulk tissue-based models (OR < 1; Fig. 1A), whereas protein-level analyses and functional assays supported a pro-tumorigenic role of COL4A6 in the CAF-stromal compartment. We attribute this bulk-compartment discordance to two factors. First, bulk RNA-seq signals are influenced by cellular composition and tumor purity; epithelial silencing/loss of COL4A6 can bias bulk-level associations. Second, mRNA abundance does not necessarily correlate with protein deposition in the ECM; stromal COL4A6 protein can accumulate despite reduced bulk mRNA signals. Therefore, our results provide an experimental framework to reconcile bulk transcriptomic associations with compartment-specific localization and function in GC. Through spatial localization (Fig. 3C) and cell-specific functional experiments (Figs. 4, 5, 6), our study provides, for the first time in GC, an experimental framework to reconcile bulk transcriptomic associations with compartment-specific localization and function: the function of COL4A6 is highly dependent on its cellular compartment. Our data support that COL4A6 is predominantly localized to the CAF/stromal compartment in situ and that COL4A6 from CAF-enriched fibroblast cultures is sufficient to enhance GC invasion, EMT, and stromal remodeling in our models. This finding is consistent with pro-tumorigenic reports in other stroma-dominated cancers like ovarian cancer and underscores the critical importance of considering cellular heterogeneity when interpreting bulk RNA-seq data.

The second core finding is the elucidation of the pro-tumorigenic mechanism of CAF-compartment-associated COL4A6. We demonstrated that COL4A6 has multifaceted roles:

• Paracrine effects on cancer cells: Through the co-culture system, we confirmed that COL4A6 secreted by CAFs (or other downstream factors) can effectively induce EMT in GC cells (E-cadherin downregulation, N-cadherin/Vimentin upregulation), thereby conferring enhanced migration and invasion capabilities (Fig. 4).

• Autocrine effects on CAFs themselves: More importantly, COL4A6 is a key inducer of CAF activation. The data in Fig. 6 clearly show that COL4A6 induces a significant phenotypic switch in CAFs. It not only upregulates the activation markers α-SMA and FAP but also reorganizes the cytoskeleton to form functional "stress fibers." This activated myofibroblast phenotype^36,37 endows CAFs with stronger matrix-remodeling capabilities (secreting and depositing collagen, Fig. 5B) and increased mechanical tension, thereby collectively constructing a "pro-tumorigenic niche" that facilitates cancer cell invasion^10,11.

The third finding of this study is the definition of clinically instructive ECM subtypes. Our macroscopic analysis confirmed that the "high-matrix (MC1)" phenotype is associated with poor prognosis, immunosuppression, and chemotherapy resistance. Through TME deconvolution, we revealed that the MC1 subtype represents a “cold tumor” enriched with CAFs, M2 macrophages, and resting dendritic cells (Supplementary Fig. 4). This provides a novel biomarker for clinical patient stratification. The characteristics of the MC1 subtype (high matrix, high CAF infiltration) are fully consistent with the oncogenic function of the CAF-COL4A6 axis we identified. Specifically, CAF-compartment-associated COL4A6, by inducing dense matrix deposition, may constitute a physical barrier that hinders the infiltration of cytotoxic T cells (resulting in a ‘cold tumor’ phenotype) and limits the penetration of chemotherapeutic agents. Our analysis indicates that the MC1 subtype is enriched in M2 macrophages and exhibits resistance to oxaliplatin, strongly suggesting that targeting the CAF-COL4A6 axis represents a promising novel strategy not only to inhibit tumor invasion but also to overcome TME immunosuppression and chemotherapy resistance. Conversely, the MC2 subtype (low-matrix/“hot tumor”) is enriched in DNA repair pathways and shows greater sensitivity to oxaliplatin (Supplementary Fig. 5), providing a rationale for precision chemotherapy.

Although this study yielded significant insights, limitations remain. First, while we clarified the function of COL4A6, the functions of the other core genes in the risk model (BGN, LAMC2) were not experimentally validated. Second, we revealed that COL4A6 induces CAF activation, but the precise upstream (what induces CAFs to secrete COL4A6) and downstream (the specific receptors, such as DDR1 or integrins, that COL4A6 relies on to activate CAFs) signaling pathways require further investigation. Third, although we used early-passage CAF-enriched fibroblast cultures and supported a CAF-like phenotype by FAP/α-SMA expression, negative lineage markers and donor-to-donor heterogeneity were not systematically assessed, which should be addressed in future studies.

Conclusion

In summary, by integrating multi-omic analysis and cell-specific functional experiments, this study systematically reconciled the “protective” bulk-transcriptome association of COL4A6 with its pro-tumorigenic, CAF-compartment-enriched function in gastric cancer. We demonstrated for the first time that while COL4A6 may exhibit a “protective” signal at the bulk-transcriptome level (potentially stemming from epithelial compartment expression loss), it is the significantly upregulated COL4A6 derived from the CAF-stromal compartment that acts as the key pro-tumorigenic factor driving GC stromal remodeling, CAF activation, and tumor progression. Our findings not only provide novel biomarkers (ECM subtypes) for precision subtyping and prognosis in GC but, more importantly, reveal the CAF-COL4A6 axis as a highly promising therapeutic target for overcoming immunosuppression and chemotherapy resistance.

Author contributions

**Qiang Sun:** Methodology, Software, Formal Analysis, Data Curation, Writing—Original Draft, Visualization.—**Shuxun Wei:** Methodology, Software, Formal Analysis, Data Curation, Writing—Original Draft, Visualization.—**Jun Yao:** Methodology, Validation, Investigation, Resources, Writing—Original Draft.—**Fuqin Wang:** Methodology, Validation, Investigation, Resources, Writing—Original Draft.—**Han Wu:** Validation, Investigation.—**Zunqi Hu:** Validation, Investigation.—**Xin Zhang:** Validation, Investigation, Data Curation.—**Dejun Yang:** Validation, Investigation, Data Curation.—**Weijun Wang:** Conceptualization, Resources, Supervision, Project Administration.—**Kai Xu:** Conceptualization, Resources, Supervision, Project Administration. All authors contributed to the Writing—Review & Editing and have read and approved the final version of the manuscript.

Funding

The authors declare that no funds, grants, or other support were received during the preparation of this manuscript.

Data availability

The datasets used and analysed during the current study available from the corresponding author on reasonable request.

Declarations

Competing interests

The authors declare no competing interests.

Ethical approval and informed consent

The authors are accountable for all aspects of the work, ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. This study was conducted in accordance with the Declaration of Helsinki (as revised in 2013). This study was approved by the Medical Ethics Committee of Shanghai Changzheng Hospital (2022SL024), and written informed consent was obtained from all patients.